Synonym Expansion Cannot Lower Recall. That Is Exactly Why Recall Is the Wrong Thing to Tune It On

说明与用途

同义词扩展同时提高召回率和精确率，驳斥了它只提高召回率的常见观点。HyDE伪答案因填充词导致效果差。

问题

Evaluating query rewriting for Retrieval-Augmented Generation (RAG) pipelines revealed that synonym expansion improves both recall and precision, contradicting the common belief that it only raises recall while lowering precision.

结果

Synonym expansion improves both recall and precision, challenging the assumption that it only raises recall. HyDE pseudo‑answers perform poorly due to filler terms. Recall is a misleading metric for additive rewrites because it cannot decrease.

工作流

• Build BM25 index over real corpus
• Apply query rewrites: verbatim, raw, expand, expandNoisy, hyde, decompose
• Score each rewrite against hand‑labelled relevance
• Compute recall@3, precision@3, MRR, mean query length
• Compare metrics across rewrites
• Analyze effect of synonym quality and query length
• Verify results with dense term‑document matrix
• Document findings

证据

Expansion raised both — recall 0.688 → 0.938 and precision 0.375 → 0.542.

来源

HyDE lost to doing nothing. MRR 0.938 → 0.323.

来源

A superset of the query can only retrieve a superset of the documents. Recall cannot fall.

来源